A novel deep transfer learning based computerized diagnostic Systems for Multi-class imbalanced diabetic retinopathy severity classification

Diabetic Retinopathy (DR) is a retinal condition that leads to gradual degeneration of the retina and eventual blindness, so early detection and evaluation of disease development are essential for effective treatment. Retinography is the most prevalent method for diagnosing DR; nevertheless, manual diagnosis is time-consuming and unpleasant. Deep Learning (DL) based algorithms have shown potential as a diagnostic tool for DR, achieving performance comparable to human image evaluation. The goal of this study is to develop Deep Transfer Learning-based Computerized Diagnostic Systems (DTL-CDS) for Multiclass DR Severity Classification (MCDR) by modifying and comparing deep inductive transfer learning (DITL) models (Inception V3, ResNet34, EfficientNet B0, VGG16, Xception). The four main objectives are i) pre-processing and balancing the imbalanced data labels, ii) proposing modified DITL model to extract features using global average pooling layer to avoid overfitting and reduce losses using leaky ReLU. The final classification uses a softmax layer to automatically classify Diabetic Retinopathy severity stages using IDRiD dataset, (iii) comparing different base and modified DITL models to existing work, and (iv) evaluating the robustness of proposed model using various performance metrics. Comprehensive analysis of MCDR between the proposed and the existing work shows that the proposed approach outperforms the state-of-the-art, with 99.36% accuracy, 0.986 precision, 0.986 recall, 0.986 F1-score, 0.997 AUC-ROC, 0.9902 AUC. The experimental results show the enhancement in diagnosis performance and modified DITL results in robust and reliable computer-aided diagnosis systems to aid specialists in proper detection of DR severity stages by reducing human errors and reducing costs.